GPT‑5.5 cuts hallucinations 52.5%

ChatGPT model updates usually sound abstract. But this one hits a very ordinary pain point — the bot confidently making things up when you ask about medicine, law, finance, or just some annoying real-world detail. OpenAI says GPT‑5.5 Instant cuts hallucinated claims by 52.5% versus GPT‑5.3 Instant on high-stakes prompts, and it is now replacing the older Instant model as ChatGPT’s default for logged-in users. ### What actually changed? The concrete change is simple: GPT‑5.5 Instant is now the default fast model inside ChatGPT, replacing GPT‑5.3 Instant. OpenAI is framing it as the everyday workhorse — the model you hit for normal questions, how-tos, translation, technical writing, image understanding, and a lot of web-backed info seeking. The company also says ChatGPT’s single “auto-switching” setup now routes requests across models more intelligently, so users see one product while the system decides what level of model is worth spending. (openai.com) ### Why is the 52.5% number a big deal? Because “hallucination” is the failure that breaks trust fastest. OpenAI’s claim is not that the model is perfect. It is that, in internal evaluations on high-stakes prompts spanning medicine, law, and finance, GPT‑5.5 Instant produced 52.5% fewer hallucinated claims than GPT‑5.3 Instant. OpenAI also says inaccurate claims fell 37.3% in especially challenging conversations users had flagged for factual errors. That second number matters because it points at messy real chats, not just neat benchmark questions. (help.openai.com) ### What is “Instant” supposed to be? Basically, the fast lane. GPT‑5.5 itself is the broader model family OpenAI launched in late April for harder work like coding, research, data analysis, and tool use. “Instant” is the cheaper, quicker version meant to answer fast without feeling flimsy. The pitch is that you should not have to choose between speed and usefulness for routine tasks anymore. OpenAI also says GPT‑5.5 matches GPT‑5.4’s per-token latency while improving capability, which helps explain why the company is comfortable making the newer Instant model the default experience. (openai.com) ### Where does the 1M-token context fit? That is the “bring the whole filing cabinet” feature. GPT‑5.5 supports a 1M-token context window, which means the model can take in very large documents, codebases, or long-running conversations in one session. The catch is pricing — OpenAI says prompts above 272K input tokens are billed at higher rates in the API. So the giant context is real, but it is most useful for developers and heavy-duty workflows, not every casual chat. (openai.com) ### Why mention coding and agents in the same breath? Because OpenAI is no longer selling these models as just answer machines. The GPT‑5.5 launch pitched stronger “agentic” behavior — planning, using tools, checking work, and moving across software until a task is done. That matters even for ChatGPT users who never touch the API, because the product is increasingly built around hidden routing: easy prompts go to lighter models, harder multi-step jobs get escalated. (developers.openai.com) Better hallucination rates are part of that story, not the whole story. ### Is this just marketing spin? Some of it is inevitably self-scored — these are OpenAI’s internal evals. But the product change is real, and it is user-facing right now: the default ChatGPT model changed, the release notes explicitly tie it to accuracy, clarity, conciseness, image understanding, STEM performance, and web-search decisions, and OpenAI published a system card for GPT‑5.5 Instant with updated safeguards. That makes this less like a vague benchmark brag and more like a platform-level behavior change. (openai.com) ### So what should a normal user expect? Not magic. Fewer weird invented details. Shorter answers. Better handling of factual lookups and long context. And maybe the biggest shift — less visible model micromanagement. ChatGPT is moving toward a “just ask” interface where OpenAI quietly decides how much model to spend on your problem. When that works, the product feels simpler. When it fails, it gets harder to know which layer made the mistake. (help.openai.com) ### Bottom line? This update is really about trust. OpenAI is trying to make the default ChatGPT experience less noisy, less error-prone, and more automatic. Cutting hallucinations by half in internal tests is the headline. The deeper move is product design — one chat box, hidden routing, and a fast model that is supposed to be good enough for most of what people actually do. (openai.com) (help.openai.com)